Replat Gemini CLI on ADK (Agent Development Kit)

[NTaylorMullen]

What would you like to be added?

Replat the Gemini CLI's core model and tool orchestration on the Google Agent Development Kit (ADK) for JavaScript/TypeScript (adk-js).

This effort will be a staged, hybrid migration to ensure zero regressions in the CLI's existing behavior—including its scheduler, policy engine, hooks, and UI event contracts—while gaining the benefits of ADK's first-class orchestration, session management, and multi-agent primitives.

Why is this needed?

The current Gemini CLI runtime is heavily coupled to direct @google/genai SDK calls. Transitioning to ADK provides:

• Modular Orchestration: Unified Runner and session lifecycle management.
• Advanced Agent Topologies: Built-in support for parallel, sequential, and loop agents.
• Improved Persistence: Native memory, artifact, and session abstractions.
• Standardization: Alignment with Google's broader agent development framework.

Additional context

This is a parent epic issue to track the overall replatting effort. Detailed sub-tasks and implementation milestones will be tracked in subsequent issues.

[ved015]

/assign

[ved015]

Hi @NTaylorMullen can i work on this one if yes can you pls add the help wanted label

[dewitt]

Thanks for opening the epic, @NTaylorMullen. I can envision this landing in three discrete phases, each building on the prior.

Phase 1: ADK TS Enhancements (largely complete)

We first need to ensure the upstream ADK library can support our CLI-specific use cases. This phase is largely complete in the upstream @google/adk repository, and documented here. We added support for native event streaming to ADK TS (matching ADK Python), introduced a utility for parsing raw event streams into strictly typed objects, added helpers for ephemeral stateless sessions, and implemented explicit explicit suspension for human-in-the-loop workflows.

Phase 2: Gemini CLI Agentic Refactoring (ready to start)

Before we actually wire up ADK though, we need to refactor the Gemini CLI core itself to establish framework-agnostic interfaces for our agents, models, and event streams. This involves decoupling our legacy execution loops and migrating them to implement these new generic interfaces natively. This allows the CLI to switch between its internal legacy runner and external agent frameworks seamlessly via configuration.

Phase 3: ADK Integration (upcoming)

Once the Phase 2 interfaces are finalized and merged, we can begin the actual an ADK integration. This phase would implement the specific adapters required to bridge ADK and our new interfaces.

I have/had a working version of this working end to end, but that code has drifted a bit since I prototyped it a week ago. But at least was a convincing demonstration of the viability of this effort, and already informed the ADK TS improvements listed above.

The hard part of course comes next though: even once this works end to end, it's critical to ensure that quality remains high, and that will require rigorous end to end testing and a deep exploration with the existing eval pipeline to identify and close any remaining gaps.

CC @abhipatel12